Core responsibilities
- Choose an architecture and training protocol
- Diagnose instability or overfitting
- Optimise the model for its deployment target
Technical skills
- PyTorch, gradients and optimisation
- Distributed training or mixed precision
- Profiling, quantisation and export
Behavioural skills
- Formulate testable hypotheses
- Keep an experiment log the team can use
Typical deliverable
A compact network that retains sensitivity to rare classes after quantisation.
Possible assessment methods
Interpret loss curves, diagnose a GPU bottleneck and propose a compute-limited ablation.
